Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A study of four instruction-tuned small language models (Llama-3.2-1B, Qwen2.5-1.5B, SmolLM2-1.7B and OLMo-2-1B) found that depth pruning collapses function calling and mathematical reasoning in every model, with retenti…
A new arXiv paper (2611.00960v1) presents Self-Refining Prompt Induction (SRPI), an offline method that turns text-to-SQL generation failures into reusable dialect instructions while keeping model weights fixed, evaluate…
Researchers released Labadain-MedQA, a Tetun medical question-answering dataset of 39,077 QA pairs created by translating MedAlpaca's Medical Flashcards and WikiDoc Patient Information collections with Gemini 2.5 Flash, …
A study posted as arXiv:2611.00983v1 finds that tokenizer segmentation, not token-level confidence alone, determines where early-exit language models spend computation inside a word, with confidence stopping giving conti…
A study of 20 large language models generating 44,782 stories in four languages found that a fine-tuning-free method called speculative naming cut character-name concentration by 95% on average, according to the arXiv pa…
A framework called CuriousTogether, introduced in arXiv paper 2611.01301v1, found that curiosity-guided multi-agent LLM interactions improved performance over non-curious settings on seven of eight tasks spanning reasoni…
An empirical study of 54 self-modification experiments across six language models found that whether and when a model stops rewriting its own system prompt is tied to model identity rather than the optimization mode assi…
Researchers released Step-dLLM, a sparse-attention method that speeds up diffusion large language model inference by periodically refreshing a block-sparse mask at anchor steps along the denoising trajectory and applying…
Keyword retrieval introduced untestable or measurement-disconfirmed criteria in 46.8% of GPT-4.1-mini ECG reports across 1,083 human-validated PTB-XL patient records, according to an arXiv study (2611.01409v1) evaluating…
A new arXiv paper (arXiv:2611.01427v1) proposes a discrete-class text provenance detection framework that separates surface-level AI polishing from AI-driven changes to discourse structure, aiming to give finer-grained s…
A new arXiv paper (2611.01378v1) introduces an information-theoretic framework for designing sparse autoencoder (SAE)-based activation steering in large language models, using Jensen-Shannon divergence (JSD) between empi…
A paper posted to arXiv (2611.01415v1) proposes Latent Harness, a method that distills an external agent harness or process-level behavior instruction into a rank-8 activation shift applied at the output of every decoder…
A large-scale trace analysis of 3.7 months of production coding-agent traffic from FreeInference found that cached input tokens account for 97.4% of input-token volume and 72.9% of modeled API cost, even though the media…
OpenAI's published API specification now lists "gpt-6-rosalind" as a possible value on the BetaModelIdsResponses and ModelIdsResponses enums, reachable from 4 OpenAI operations. The addition is producer-side additive but…
SemiAnalysis released GDPx v1.0, a benchmark of 206 long-running agentic tasks across 11 domains built with 14 partners, and found that the best model, Anthropic's Claude Opus 5.5, passed only 12.0% of tasks at $51.41 pe…
Anthropic's Frontier Red Team, working with EPFL, MIT and StarkWare, released FlashProve, an open-source prototype that generates a cryptographic proof that a specific language model produced a specific reply. The prover…
A study comparing replacement operators for training-free MoE expert pruning, holding the pruning mask fixed, evaluated three operator families on OLMoE, DeepSeek-V2-Lite and Mo (a third model whose name is truncated in …
A developer released FreeLLMAPI, a self-hosted server that aggregates the free tiers of 34 LLM providers behind a single OpenAI-compatible /v1 API, with Anthropic wire-format support so Claude Code and the Anthropic SDKs…
Mistral released the open weights for Mistral Large 4, a 1.05-trillion-parameter multimodal Mixture-of-Experts model nicknamed "Le Chonk," on October 29, with checkpoints in FP8, NVFP4 and GGUF formats live on HuggingFac…
A large-scale study using PaperTrajectory-2026, the 2026 release of the PaperTrajectory dataset covering 6,959 paper trajectories, investigates whether AI reviewers' judgments reflect the improvements authors make throug…